Member-only story
Data Preprocessing for Deep Learning
One-hot Encode Scalar-value Labels for Deep Learning Models
One-hot vector explained in plain English
We need to convert scalar-value labels into a one-hot vector before using them in deep learning models.
This is required for multiclass classification models that output probabilities per class when using the categorical_crossentropy loss function.
Sparse scalar representation
The values in the label column are usually represented as sparse scalars (i.e. in single-digit format). For example, the training and test labels in the MNIST digits dataset are represented in single-digit format ranging from 0 to 9. Each digit represents a class label.
The training and test labels are in two separate one-dimensional vectors.
One-hot representation
As I explained at the beginning of this article, sparse scalar representation is not suitable for multiclass classification models that output probabilities per class. So, it is necessary to perform one-hot encoding for scalar-value labels before using them in deep learning models.
